------------------------------------------------------------------------------
Cambridge synthetic light field dataset for testing view interpolation methods
------------------------------------------------------------------------------

This dataset contains the synthetic light field dataset reported in the paper: 

A Benchmark of Light Field View Interpolation Methods
Dingcheng Yue, Muhammad Shahzeb Khan Gul, Michel Bätz, Joachim Keinert, and Rafał Mantiuk
In: 2020 IEEE International Conference on Multimedia & Expo Workshops (ICMEW), 1–6. 
https://doi.org/10.1109/ICMEW46912.2020.9106041

If you use this dataset in your work, please cite the paper above. 


Organization of the dataset
---------------------------

The dataset is organised as follows:

* "scene.csv": a CSV file specifies what each scene is conditioned on.
* "<scene>/<frame_number>/<mean_disparity>/<row>_<column>.png": a 512x512 image
  file contains a view of a specific light field.
* "<scene>/<scene>.blend": the original blender file that generates the scene.
* "<scene>/<frame_number>/disparity.npy": a numpy array specifying the disparity map of
  the central view of the light field. The disparity is computed assuming that the
  baseline is 1 unit.
* "<scene>/<frame_number>/<mean_disparity>/param.txt" specify the parameter for
  the extrinsic camera matrix. Specifically, "num_x", "num_y" specifies how many rows and
  columns of views are there in the light field. And "base_x" and "base_y"
  specifies the distance between cameras of two neighbouring views.

The directory and file names represent the following information:

+ <scene> - ID of a scene (content). Note that each scene is varied by a specific
          condition (disparity, occlusion, non-lambertian) showed in scene.csv
+ <frame_number> - The frame number associated with each scene. For the scenes that vary
                 by disparity, there is only one frame number (001). For the scenes that
                 vary in occlusion and non-Lambertian surfaces, the <frame_number>
                 varies between (001) and (010)
+ <mean_disparity> - The mean disparity between two neighbouring views of the
                   scene.
+ <row> - The light field consists of a set of 9x9 views. The row number of a
        particular view. (y-direction)
+ <column> - The light field consists of a set of 9x9 views. The columns number
            of a particular view. (x direction)


Camera Parameters
------------------

All the scenes use the same intrinsic camera parameters, specifically, the field of view is 39.6 degrees, the aspect ratio is 1 (square 512x512px images), the near and far clipping planes are at 0.1 and 100. A symmetric view frustum is used. 

The intrinsic camera matrix is:

Matrix(((2.777777671813965, 0.0, 0.0, 0.0),
        (0.0, 2.777777671813965, 0.0, 0.0),
        (0.0, 0.0, -1.0020020008087158, -0.20020020008087158),
        (0.0, 0.0, -1.0, 0.0)))

The extrinsic matrix can be calculated by base_x and base_y specified in the param.txt

Copyright
---------

Refer to the copyright.txt for the copyright information on images used in the dataset.
All blender files are licensed under a CC license.


Questions
---------

Any questions regarding the dataset should be directed to:
Rafal Mantiuk <rkm38@cam.ac.uk> and Dingcheng Yue <dy276@cam.ac.uk>
(include both authors in the correspondence).